19 |
* along with this program ; if not, write to the Free Software |
* along with this program ; if not, write to the Free Software |
20 |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
21 |
* |
* |
22 |
* $Id: xvid.c,v 1.49 2004-04-01 11:11:28 suxen_drol Exp $ |
* $Id: xvid.c,v 1.56 2004-10-17 10:20:15 edgomez Exp $ |
23 |
* |
* |
24 |
****************************************************************************/ |
****************************************************************************/ |
25 |
|
|
52 |
unsigned int xvid_debug = 0; /* xvid debug mask */ |
unsigned int xvid_debug = 0; /* xvid debug mask */ |
53 |
#endif |
#endif |
54 |
|
|
55 |
#if defined(ARCH_IS_IA32) |
#if defined(ARCH_IS_IA32) && defined(_MSC_VER) |
|
#if defined(_MSC_VER) |
|
56 |
# include <windows.h> |
# include <windows.h> |
57 |
#else |
#elif defined(ARCH_IS_IA32) || defined(ARCH_IS_PPC) |
58 |
# include <signal.h> |
# include <signal.h> |
59 |
# include <setjmp.h> |
# include <setjmp.h> |
60 |
|
|
77 |
* 0 : SIGILL was *not* signalled |
* 0 : SIGILL was *not* signalled |
78 |
* 1 : SIGILL was signalled |
* 1 : SIGILL was signalled |
79 |
*/ |
*/ |
80 |
|
#if defined(ARCH_IS_IA32) && defined(_MSC_VER) |
81 |
int |
static int |
82 |
sigill_check(void (*func)()) |
sigill_check(void (*func)()) |
83 |
{ |
{ |
|
#if defined(_MSC_VER) |
|
84 |
_try { |
_try { |
85 |
func(); |
func(); |
86 |
} |
} _except(EXCEPTION_EXECUTE_HANDLER) { |
|
_except(EXCEPTION_EXECUTE_HANDLER) { |
|
87 |
|
|
88 |
if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION) |
if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION) |
89 |
return 1; |
return(1); |
90 |
} |
} |
91 |
return 0; |
return(0); |
92 |
#else |
} |
93 |
|
#elif defined(ARCH_IS_IA32) || defined(ARCH_IS_PPC) |
94 |
|
static int |
95 |
|
sigill_check(void (*func)()) |
96 |
|
{ |
97 |
void * old_handler; |
void * old_handler; |
98 |
int jmpret; |
int jmpret; |
99 |
|
|
100 |
|
/* Set our SIGILL handler */ |
101 |
old_handler = signal(SIGILL, sigill_handler); |
old_handler = signal(SIGILL, sigill_handler); |
102 |
if (old_handler == SIG_ERR) |
|
103 |
{ |
/* Check for error */ |
104 |
return -1; |
if (old_handler == SIG_ERR) { |
105 |
|
return(-1); |
106 |
} |
} |
107 |
|
|
108 |
|
/* Save stack context, so if func triggers a SIGILL, we can still roll |
109 |
|
* back to a valid CPU state */ |
110 |
jmpret = setjmp(mark); |
jmpret = setjmp(mark); |
111 |
if (jmpret == 0) |
|
112 |
{ |
/* If setjmp returned directly, then its returned value is 0, and we still |
113 |
|
* have to test the passed func. Otherwise it means the stack context has |
114 |
|
* been restored by a longjmp() call, which in our case happens only in the |
115 |
|
* signal handler */ |
116 |
|
if (jmpret == 0) { |
117 |
func(); |
func(); |
118 |
} |
} |
119 |
|
|
120 |
|
/* Restore old signal handler */ |
121 |
signal(SIGILL, old_handler); |
signal(SIGILL, old_handler); |
122 |
|
|
123 |
return jmpret; |
return(jmpret); |
|
#endif |
|
124 |
} |
} |
125 |
#endif |
#endif |
126 |
|
|
142 |
#endif |
#endif |
143 |
|
|
144 |
#if defined(ARCH_IS_PPC) |
#if defined(ARCH_IS_PPC) |
145 |
#if defined(ARCH_IS_PPC_ALTIVEC) |
if (!sigill_check(altivec_trigger)) |
146 |
cpu_flags |= XVID_CPU_ALTIVEC; |
cpu_flags |= XVID_CPU_ALTIVEC; |
147 |
#endif |
#endif |
|
#endif |
|
148 |
|
|
149 |
return cpu_flags; |
return cpu_flags; |
150 |
} |
} |
222 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_c; |
223 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_c; |
224 |
|
|
225 |
|
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_c; |
226 |
|
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_c; |
227 |
|
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_c; |
228 |
|
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_c; |
229 |
|
|
230 |
interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c; |
interpolate16x16_lowpass_h = interpolate16x16_lowpass_h_c; |
231 |
interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c; |
interpolate16x16_lowpass_v = interpolate16x16_lowpass_v_c; |
232 |
interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c; |
interpolate16x16_lowpass_hv = interpolate16x16_lowpass_hv_c; |
307 |
dev16 = dev16_c; |
dev16 = dev16_c; |
308 |
sad16v = sad16v_c; |
sad16v = sad16v_c; |
309 |
sse8_16bit = sse8_16bit_c; |
sse8_16bit = sse8_16bit_c; |
310 |
|
sse8_8bit = sse8_8bit_c; |
311 |
|
|
312 |
#if defined(ARCH_IS_IA32) |
#if defined(ARCH_IS_IA32) |
313 |
|
|
362 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_mmx; |
363 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_mmx; |
364 |
|
|
365 |
|
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_mmx; |
366 |
|
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_mmx; |
367 |
|
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_mmx; |
368 |
|
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_mmx; |
369 |
|
|
370 |
interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx; |
interpolate8x8_6tap_lowpass_h = interpolate8x8_6tap_lowpass_h_mmx; |
371 |
interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx; |
interpolate8x8_6tap_lowpass_v = interpolate8x8_6tap_lowpass_v_mmx; |
372 |
|
|
408 |
dev16 = dev16_mmx; |
dev16 = dev16_mmx; |
409 |
sad16v = sad16v_mmx; |
sad16v = sad16v_mmx; |
410 |
sse8_16bit = sse8_16bit_mmx; |
sse8_16bit = sse8_16bit_mmx; |
411 |
|
sse8_8bit = sse8_8bit_mmx; |
412 |
} |
} |
413 |
|
|
414 |
/* these 3dnow functions are faster than mmx, but slower than xmm. */ |
/* these 3dnow functions are faster than mmx, but slower than xmm. */ |
436 |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_xmm; |
437 |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_xmm; |
438 |
|
|
439 |
|
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_xmm; |
440 |
|
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_xmm; |
441 |
|
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_xmm; |
442 |
|
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_xmm; |
443 |
|
|
444 |
/* reduced resolution */ |
/* reduced resolution */ |
445 |
copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_xmm; |
copy_upsampled_8x8_16to8 = xvid_Copy_Upsampled_8x8_16To8_xmm; |
446 |
add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_xmm; |
add_upsampled_8x8_16to8 = xvid_Add_Upsampled_8x8_16To8_xmm; |
479 |
|
|
480 |
if ((cpu_flags & XVID_CPU_3DNOWEXT)) { |
if ((cpu_flags & XVID_CPU_3DNOWEXT)) { |
481 |
|
|
|
/* Inverse DCT */ |
|
|
idct = idct_3dne; |
|
|
|
|
482 |
/* Buffer transfer */ |
/* Buffer transfer */ |
483 |
transfer_8to16copy = transfer_8to16copy_3dne; |
transfer_8to16copy = transfer_8to16copy_3dne; |
484 |
transfer_16to8copy = transfer_16to8copy_3dne; |
transfer_16to8copy = transfer_16to8copy_3dne; |
485 |
transfer_8to16sub = transfer_8to16sub_3dne; |
transfer_8to16sub = transfer_8to16sub_3dne; |
486 |
transfer_8to16subro = transfer_8to16subro_3dne; |
transfer_8to16subro = transfer_8to16subro_3dne; |
|
transfer_8to16sub2 = transfer_8to16sub2_3dne; |
|
487 |
transfer_16to8add = transfer_16to8add_3dne; |
transfer_16to8add = transfer_16to8add_3dne; |
488 |
transfer8x8_copy = transfer8x8_copy_3dne; |
transfer8x8_copy = transfer8x8_copy_3dne; |
489 |
|
|
490 |
|
if ((cpu_flags & XVID_CPU_MMXEXT)) { |
491 |
|
/* Inverse DCT */ |
492 |
|
idct = idct_3dne; |
493 |
|
|
494 |
|
/* Buffer transfer */ |
495 |
|
transfer_8to16sub2 = transfer_8to16sub2_3dne; |
496 |
|
|
497 |
|
/* Interpolation */ |
498 |
|
interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne; |
499 |
|
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne; |
500 |
|
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne; |
501 |
|
|
502 |
/* Quantization */ |
/* Quantization */ |
503 |
quant_h263_intra = quant_h263_intra_3dne; |
quant_h263_intra = quant_h263_intra_3dne; /* cmov only */ |
504 |
quant_h263_inter = quant_h263_inter_3dne; |
quant_h263_inter = quant_h263_inter_3dne; |
505 |
dequant_mpeg_intra = dequant_mpeg_intra_3dne; |
dequant_mpeg_intra = dequant_mpeg_intra_3dne; /* cmov only */ |
506 |
dequant_mpeg_inter = dequant_mpeg_inter_3dne; |
dequant_mpeg_inter = dequant_mpeg_inter_3dne; |
507 |
dequant_h263_intra = dequant_h263_intra_3dne; |
dequant_h263_intra = dequant_h263_intra_3dne; |
508 |
dequant_h263_inter = dequant_h263_inter_3dne; |
dequant_h263_inter = dequant_h263_inter_3dne; |
509 |
|
|
510 |
/* ME functions */ |
/* ME functions */ |
511 |
calc_cbp = calc_cbp_3dne; |
calc_cbp = calc_cbp_3dne; |
512 |
|
|
513 |
sad16 = sad16_3dne; |
sad16 = sad16_3dne; |
514 |
sad8 = sad8_3dne; |
sad8 = sad8_3dne; |
515 |
sad16bi = sad16bi_3dne; |
sad16bi = sad16bi_3dne; |
516 |
sad8bi = sad8bi_3dne; |
sad8bi = sad8bi_3dne; |
517 |
dev16 = dev16_3dne; |
dev16 = dev16_3dne; |
518 |
|
} |
|
/* Interpolation */ |
|
|
interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_3dne; |
|
|
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_3dne; |
|
|
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_3dne; |
|
519 |
} |
} |
520 |
|
|
521 |
if ((cpu_flags & XVID_CPU_SSE2)) { |
if ((cpu_flags & XVID_CPU_SSE2)) { |
535 |
/* DCT operators |
/* DCT operators |
536 |
* no iDCT because it's not "Walken matching" */ |
* no iDCT because it's not "Walken matching" */ |
537 |
fdct = fdct_sse2_skal; |
fdct = fdct_sse2_skal; |
538 |
|
|
539 |
|
/* postprocessing */ |
540 |
|
image_brightness = image_brightness_sse2; |
541 |
} |
} |
542 |
#endif /* ARCH_IS_IA32 */ |
#endif /* ARCH_IS_IA32 */ |
543 |
|
|
568 |
#endif |
#endif |
569 |
|
|
570 |
#if defined(ARCH_IS_PPC) |
#if defined(ARCH_IS_PPC) |
571 |
if ((cpu_flags & XVID_CPU_ASM)) |
if ((cpu_flags & XVID_CPU_ALTIVEC)) { |
572 |
{ |
/* sad operators */ |
573 |
calc_cbp = calc_cbp_ppc; |
sad16 = sad16_altivec_c; |
574 |
} |
sad16bi = sad16bi_altivec_c; |
575 |
|
sad8 = sad8_altivec_c; |
576 |
|
dev16 = dev16_altivec_c; |
577 |
|
|
578 |
|
sse8_16bit = sse8_16bit_altivec_c; |
579 |
|
|
580 |
|
/* mem transfer */ |
581 |
|
transfer_8to16copy = transfer_8to16copy_altivec_c; |
582 |
|
transfer_16to8copy = transfer_16to8copy_altivec_c; |
583 |
|
transfer_8to16sub = transfer_8to16sub_altivec_c; |
584 |
|
transfer_8to16subro = transfer_8to16subro_altivec_c; |
585 |
|
transfer_8to16sub2 = transfer_8to16sub2_altivec_c; |
586 |
|
transfer_16to8add = transfer_16to8add_altivec_c; |
587 |
|
transfer8x8_copy = transfer8x8_copy_altivec_c; |
588 |
|
|
589 |
if ((cpu_flags & XVID_CPU_ALTIVEC)) |
/* Inverse DCT */ |
590 |
{ |
idct = idct_altivec_c; |
591 |
calc_cbp = calc_cbp_altivec; |
|
592 |
fdct = fdct_altivec; |
/* Interpolation */ |
593 |
idct = idct_altivec; |
interpolate8x8_halfpel_h = interpolate8x8_halfpel_h_altivec_c; |
594 |
sadInit = sadInit_altivec; |
interpolate8x8_halfpel_v = interpolate8x8_halfpel_v_altivec_c; |
595 |
sad16 = sad16_altivec; |
interpolate8x8_halfpel_hv = interpolate8x8_halfpel_hv_altivec_c; |
596 |
sad8 = sad8_altivec; |
|
597 |
dev16 = dev16_altivec; |
interpolate8x8_avg2 = interpolate8x8_avg2_altivec_c; |
598 |
|
interpolate8x8_avg4 = interpolate8x8_avg4_altivec_c; |
599 |
|
|
600 |
|
interpolate8x8_halfpel_add = interpolate8x8_halfpel_add_altivec_c; |
601 |
|
interpolate8x8_halfpel_h_add = interpolate8x8_halfpel_h_add_altivec_c; |
602 |
|
interpolate8x8_halfpel_v_add = interpolate8x8_halfpel_v_add_altivec_c; |
603 |
|
interpolate8x8_halfpel_hv_add = interpolate8x8_halfpel_hv_add_altivec_c; |
604 |
|
|
605 |
|
/* Colorspace conversion */ |
606 |
|
bgra_to_yv12 = bgra_to_yv12_altivec_c; |
607 |
|
abgr_to_yv12 = abgr_to_yv12_altivec_c; |
608 |
|
rgba_to_yv12 = rgba_to_yv12_altivec_c; |
609 |
|
argb_to_yv12 = argb_to_yv12_altivec_c; |
610 |
|
|
611 |
|
yuyv_to_yv12 = yuyv_to_yv12_altivec_c; |
612 |
|
uyvy_to_yv12 = uyvy_to_yv12_altivec_c; |
613 |
|
|
614 |
|
yv12_to_yuyv = yv12_to_yuyv_altivec_c; |
615 |
|
yv12_to_uyvy = yv12_to_uyvy_altivec_c; |
616 |
|
|
617 |
|
/* Quantization */ |
618 |
|
quant_h263_intra = quant_h263_intra_altivec_c; |
619 |
|
quant_h263_inter = quant_h263_inter_altivec_c; |
620 |
|
dequant_h263_intra = dequant_h263_intra_altivec_c; |
621 |
|
dequant_h263_inter = dequant_h263_inter_altivec_c; |
622 |
|
|
623 |
|
/* Qpel stuff */ |
624 |
|
xvid_QP_Funcs = &xvid_QP_Funcs_Altivec_C; |
625 |
|
xvid_QP_Add_Funcs = &xvid_QP_Add_Funcs_Altivec_C; |
626 |
} |
} |
627 |
#endif |
#endif |
628 |
|
|
630 |
xvid_debug = init->debug; |
xvid_debug = init->debug; |
631 |
#endif |
#endif |
632 |
|
|
633 |
return 0; |
return(0); |
634 |
} |
} |
635 |
|
|
636 |
|
|
641 |
return XVID_ERR_VERSION; |
return XVID_ERR_VERSION; |
642 |
|
|
643 |
info->actual_version = XVID_VERSION; |
info->actual_version = XVID_VERSION; |
644 |
info->build = "xvid-1.0.0"; |
info->build = "xvid-1.1-cvshead"; |
645 |
info->cpu_flags = detect_cpu_flags(); |
info->cpu_flags = detect_cpu_flags(); |
646 |
|
|
647 |
#if defined(_SMP) && defined(WIN32) |
#if defined(_SMP) && defined(WIN32) |